A shared map for AI in education
In October 2019 I wrote and presented a report for UNESCO, proposing how a global platform for AI in education might work. This is what it said.
The Beijing Consensus on AI and education, agreed in May 2019, asked UNESCO to set up an “AI for Education” platform: a clearinghouse for AI courses, tools, policies and good practice. I was invited to write a concept proposal for it, for Fengchun Miao’s team. The work was unpaid. I presented it at UNESCO in Paris in October 2019.
The report was called Global AI Architecture for Education (GAIA 4Ed). Its central idea was that the hard part is not the AI. It is getting the world’s education systems to describe their problems and their data in ways that can be compared, without forcing everyone into one mould.
The presentation led to a verbal offer of a contract. I declined it, because I was moving to New Zealand.
Two drafts survive; the version I presented does not.
What the report proposed
Two kinds of user, one system. Policy makers need something like a business-intelligence view: progress towards the education Sustainable Development Goal by place and over time, and a way to model what an intervention might do if scaled up. Practitioners — teachers, schools, learners, the AI community — need shared tools, content and models of teaching. Each should be able to see the other’s side.
Three groups of products.
- A policy portal: a dashboard of progress, open data with context, and impact modelling.
- A practitioners’ portal: a shared model of teaching, a lifelong learning profile, and common maps of subject content.
- “Hi AI!”: a free open course on human and artificial intelligence, and how to use each well.
A pyramid of knowledge. Content organised by the Consensus’s own headings, from international standards and national policies down to cases, resources, research and data. It could be read top-down or bottom-up.
Open challenges, both ways. Algorithms posted with their use cases and past results, for educators to find and apply. And the reverse: educators post a problem, with context, anonymised data and the outcome they want, and the AI community responds.
Treat problem definition as the main work. Framing a problem, describing its data and cleaning it is most of any AI project. The platform should help educators get their challenges “AI-ready” in a standard form: why, context, data, desired outcome.
Common maps, with local variety kept. Different countries, and often different bodies within one country, describe the same learning in different words and structures. The report proposed a common model for each subject. Anyone could upload their own version, matched label by label to the common one, adding detail where theirs had more. AI could suggest the matches. For the first time, it argued, you could see where the world’s teaching profession agrees and where it differs — for example, in a global map of maths.
Start off-grid. Inclusion means schools with no power, internet or devices. Every product should have a paper version — topic tests, mark schemes, a return sheet — that could be posted back for entry. The report argued this should be one of the first things built, because it pins down the real shape of the problem.
Anonymise at the point of entry. Only the school knows a pupil’s name. The platform stores results against random keys, so the same learner can be followed over time without being identified.
Build it to learn. Consult each user group early, design for feedback at every step, and let the platform’s structure evolve with use.
Looking back
Its weakness was scope. It tried to be a dashboard, a data commons, a shared vocabulary, a marketplace and a course at once. Its strongest idea was mapping between curricula, which language models now make far more practical than in 2019. That idea became the starting point for my later work on Cross-Curricular Correspondence.